Department of Preventive Medicine Biostatistics Seminar Series: Dynamic Prediction of Optimal Timing for Initiating Dialysis Using a Multistate Framework

The Division of Biostatistics of the Department of Preventive Medicine at UT Health Sciences invites you to attend the following seminar.

When: Monday, August 31, 2026, 2:00 - 3:00 pm CT

ZOOM Virtual Room Connection: Register in advance for this meeting to get the Zoom Link

Seminar Website: https://www.uthsc.edu/preventive-medicine/events.php

Dynamic Prediction of Optimal Timing for Initiating Dialysis Using a Multistate Framework
Yue Zhan
University of Nebraska Medical Center

The decision to initiate a non-reversible intervention can be challenging because it involves balancing potential survival benefits against substantial treatment burdens. For example, the decision to initiate dialysis for patients with end-stage chronic kidney disease may depend on evolving biomarkers, such as estimated glomerular filtration rate (eGFR), and other clinical symptoms that are predictive of pre-treatment survival. Standard survival analysis methods may be subject to immortal time bias, confounding by indication, and limited generalizability between treated and untreated populations. To address these potential biases, we develop a multistate framework for dynamically comparing alternative treatment-initiation strategies using longitudinal biomarker history. We consider a three-state illness-treatment-death model that incorporates both death before treatment and survival after treatment. At prespecified decision times, a linear mixed-effects model summarizes each patient's biomarker trajectory, and predicted biomarker values are incorporated into transition-specific Cox models. For patients who are alive and untreated at each decision time, potential residual survival under immediate and delayed treatment strategies is predicted using multistate g-computation and summarized by restricted residual mean survival time. The resulting survival contrast is used to guide individualized treatment recommendations. Subject-level bootstrap resampling is used for statistical inference. Simulation studies evaluate model parameter estimation, predicted survival benefits, and the accuracy of individualized treatment decisions.